Overview
anyrun is a Docker sandbox execution engine for AI Agents. It provides secure, isolated code execution with full observability — automatically collecting execution traces, discovering error patterns and success paths via statistical rules, extracting reusable experience as SKILL.md through LLM, and enabling autonomous skill evolution (beta → prod → decayed → retired).
Key features:
- Zero-config Docker sandbox: one line of Python to run any code
securely
- Full observability: every execution is traced to SQLite + JSON
- Pattern discovery: error clustering, success path detection, anomaly
detection
- Experience extraction: LLM auto-generates SKILL.md from execution
patterns
- Self-evolution: skills auto-upgrade, decay, and get repaired
- MCP native: exposes all tools (sandbox_run, file ops, toolbox
management) via MCP protocol
Use cases: AI Agent tool execution, secure code sandboxing, agent
self-improvement pipelines, CI/CD code execution.
Server Config
{
"mcpServers": {
"anyrun": {
"command": "python3",
"args": [
"-m",
"anyrun.mcp_server"
]
}
}
}